Skip to content

Latest commit

 

History

History
14 lines (12 loc) · 1.52 KB

REAL-WORLD_INFERENCE.md

File metadata and controls

14 lines (12 loc) · 1.52 KB

Inference

Seer has been encapsulated into a real-world controller, allowing for straightforward adaptation to downstream tasks. To ensure smooth implementation and avoid common errors, we offer the following recommendations:

  • 🔥 Proprio First v.s. Image First: During data collection and inference, always acquire robot proprioception data first, followed by image observations. This order minimizes the timestep interval since capturing images is significantly more time-intensive than reading proprioception data.
  • 🔥 Delta-to-Absolute Action Labels: To simplify matrix computation and transformation, we provide a delta-action-to-absolute-action conversion in the deployment script. This aligns with the absolute-action-to-delta-action transformation found in the post-process script.
  • 🔥 Consistent Control Frequency: Ensure that the control frequencies used during data collection match those during inference. Discrepancies in frequency can lead to inconsistent results.

⭐ Real-World Controller

A wrapped seer controller is provided for real-world deployment. This controller is modular and can be easily adapted to specific tasks or environments.

🌟 Real-World Deployment Pseudocode

To deploy the wrapped Seer controller for real-world tasks, modify the deployment script to fit your specific environment. Then, execute the deployment with the following command:

bash scripts/REAL/deploy.sh